Periodic Reporting for period 1 - MultiSeaSpace (Developing a unified spatial modelling strategy that accounts for interactions between species at different marine trophic levels, and different types of survey data.)
Période du rapport: 2019-11-01 au 2021-10-31
The conservation of marine ecosystems for a productive human exploitation is a challenging task. A base principle of the ecosystem based management approach is to manage resources by keeping them as healthy as possible. To do so, we require statistical modelling techniques that maximize the the amount of inferable information from the available data. Unfortunately, collecting data on species in marine ecosystems is particularly challenging as the marine environment is largely inaccessible and individuals are mostly invisible to researchers. Surveys can typically only collect information on a limited number of aspects of individuals’ distribution in space, mainly dependant on the behaviour of a specific species within space and other practical limitations. As a result, a number of different sampling methods have been used, producing different data structures that require different statistical modelling approaches.
Different statistical methods are often applied using different software or packages. This project was aimed at unifying different spatial distribution modelling approaches in a single software that allowed to:
(a) Model habitat preferences of a range of different species within the same modelling framework, independent of the specific data collection approach;
(b) jointly modelling data from different surveys on the same species;
(c) jointly modelling data from the same/different surveys on several species.
Specifically, we used the software package inlabru (www.inlabru.org/inlabru) which is based on integrated nested Laplace approximation (INLA) and the associated software R-INLA.
Different statistical methods are often applied using different software or packages. This project was aimed at unifying different spatial distribution modelling approaches in a single software that allowed to:
(a) Model habitat preferences of a range of different species within the same modelling framework, independent of the specific data collection approach;
(b) jointly modelling data from different surveys on the same species;
(c) jointly modelling data from the same/different surveys on several species.
Specifically, we used the software package inlabru (www.inlabru.org/inlabru) which is based on integrated nested Laplace approximation (INLA) and the associated software R-INLA.
MultiSeaSpace mainly focused in the following 5 developments:
1 - Inference of species interactions through spatial latent fields in inlabru: after several months of simulation based validation of the method proposed in the academic literature to infer spatial species interaction through latent fields, I did not obtain reliable results using the method. Spatial latent fields, as most latent fields, are very flexible modeling structures. After careful considerations, I concluded that the use of spatial latents fields (and potentially other latent fields not investigated here) are prone to inferring false positive interactions among species.
2 - Fishery survey data integration in inlabru: I performed a simulation study to validate the model design, which promptly demonstrated to be successfull. Then I tested three different case studies, a red mullet and a european hake case study in the Mediterranean combining trawl survey and trawl commercial catch data, and a common sole case study in the Bay of Biscay combining trawl survey and trammel net commercial catches. The first two case studies turned out to be unsuccessful due the the differences in catchability coefficients between the scientific survey gear and the commercial fishing gear. The common sole case study showed consistent patterns and trends, thus it was used to prepare a scientific manuscript that is now under review in the ICES journal of Sea Research. All the code is available in github.
3 - Analysis of telemetry data using Log-Gaussian Cox Processes in inlabru: I compared three different methods to infer species habitat use through telemetry data. The use of pseudo-absences to model the data using a binomial distribution; using a Poisson discretization of the data; and using a spatially continuous LGCP model. The latter two methods produce very similar results to that of the traditional binomial model using pseudo-absences, but without the burden of having to carefully create pseudo-absences to model the data. In addition, the binomial modelling approach is known (and we also saw this) to be affected by the disposition of the pseudo-absences in the data, which does not hapen with the Poisson and LGCP approaches. Lastly, between the LGCP and Poisson approximation, the LGCP requires less steps to do the modelling because we don't need to aggregate observations in a grid or voronoy discretization to fit the model.
4 - Accommodating fishery extreme catch events in fisheries distribution models: Fishery data is characterised, among other things, by a large number of zeros as well as sporadic extremely high catch abundances. While zero inflation has been tackled in a number of ways, extreme catches has generally been either neglected or windsorized in fishery models. Within this framework, we have proposed a spliced model that not only tackles zero inflation, but also accommodates extreme observations by means of a pareto distribution.
5 - Analysis of fishery acoustic data through Marked Point Processes in inlabru: This is pretty much work in progress. The conceptual design and simulation testing is ready for both 2D and 3D models. The modelling of real data has commenced using a 2D Marked Point Process model that we may further develop into a 3D model if the 2D model proves to be useful. The modeI expects to provide a more realistic modeling of the process behind acoustic data, and therefore a better estimation of uncertainty in abundance estimates.
1 - Inference of species interactions through spatial latent fields in inlabru: after several months of simulation based validation of the method proposed in the academic literature to infer spatial species interaction through latent fields, I did not obtain reliable results using the method. Spatial latent fields, as most latent fields, are very flexible modeling structures. After careful considerations, I concluded that the use of spatial latents fields (and potentially other latent fields not investigated here) are prone to inferring false positive interactions among species.
2 - Fishery survey data integration in inlabru: I performed a simulation study to validate the model design, which promptly demonstrated to be successfull. Then I tested three different case studies, a red mullet and a european hake case study in the Mediterranean combining trawl survey and trawl commercial catch data, and a common sole case study in the Bay of Biscay combining trawl survey and trammel net commercial catches. The first two case studies turned out to be unsuccessful due the the differences in catchability coefficients between the scientific survey gear and the commercial fishing gear. The common sole case study showed consistent patterns and trends, thus it was used to prepare a scientific manuscript that is now under review in the ICES journal of Sea Research. All the code is available in github.
3 - Analysis of telemetry data using Log-Gaussian Cox Processes in inlabru: I compared three different methods to infer species habitat use through telemetry data. The use of pseudo-absences to model the data using a binomial distribution; using a Poisson discretization of the data; and using a spatially continuous LGCP model. The latter two methods produce very similar results to that of the traditional binomial model using pseudo-absences, but without the burden of having to carefully create pseudo-absences to model the data. In addition, the binomial modelling approach is known (and we also saw this) to be affected by the disposition of the pseudo-absences in the data, which does not hapen with the Poisson and LGCP approaches. Lastly, between the LGCP and Poisson approximation, the LGCP requires less steps to do the modelling because we don't need to aggregate observations in a grid or voronoy discretization to fit the model.
4 - Accommodating fishery extreme catch events in fisheries distribution models: Fishery data is characterised, among other things, by a large number of zeros as well as sporadic extremely high catch abundances. While zero inflation has been tackled in a number of ways, extreme catches has generally been either neglected or windsorized in fishery models. Within this framework, we have proposed a spliced model that not only tackles zero inflation, but also accommodates extreme observations by means of a pareto distribution.
5 - Analysis of fishery acoustic data through Marked Point Processes in inlabru: This is pretty much work in progress. The conceptual design and simulation testing is ready for both 2D and 3D models. The modelling of real data has commenced using a 2D Marked Point Process model that we may further develop into a 3D model if the 2D model proves to be useful. The modeI expects to provide a more realistic modeling of the process behind acoustic data, and therefore a better estimation of uncertainty in abundance estimates.
The modeling methods developed in MultiSeaSpace have improved the state of the art in modeling fishery and telemetry data.
1) Data integration of fishery surveys allows us to use opportunisticaly collected data to balance unexpected gaps in scientific survey data. These gaps can be either spatial, temporal and, under certain assumptions, even spatiotemporal gaps, that generaly occur due to financial problems, bad weather, covid, etc. This research line may have a profound economic impact in the future of fishery data collection as we may be able to reduce the costs of data collection by a large amount. This may be particularly relevant in case of financial problems or the need of collecting more fishery data for management purposes.
2) The accomodation of fishery extreme catch events (ECEs) has often been a neglected issue in the fishery modeling community. In light of our results, ECEs may play an important role in the correct estimation of fish abundance. Removing or windsorizing ECEs may result in an underestimation of abundance. This may imply that certain species' stock assessments have underestimated abundance. Despite the short term negative economic impact of this finding, driven by the modeling cost of this model, it will as well have a long term positive economic impact on the fishery.
3) The analysis of telemetry data using LGCPs to infer species habitat use will reduce modelling cost as it simplifies the required steps to model this sort of data. This may as well be applied to fishery distribution data, i.e. fishing locations, to account for the preferential sampling bias problem when estimating fish abundance using fishery dependent data only.
4) The analysis of acoustic data using Marked Point Processes, despite being work in progress, is an idea that was seeded within MultiSeaSpace. This new way of modeling fishery acoustic data provides a unique and statisticaly sounf tool to provide a probabilistic estimates abundance in this surveys. It is hard to estimate the economic impact of this for us, but if it works, we expect it to be an inflexion point in the estimation of uncertainty in fishery acoustic surveys, which will probably imply better fit in stock assessment models and therefore the management of the fishery.
1) Data integration of fishery surveys allows us to use opportunisticaly collected data to balance unexpected gaps in scientific survey data. These gaps can be either spatial, temporal and, under certain assumptions, even spatiotemporal gaps, that generaly occur due to financial problems, bad weather, covid, etc. This research line may have a profound economic impact in the future of fishery data collection as we may be able to reduce the costs of data collection by a large amount. This may be particularly relevant in case of financial problems or the need of collecting more fishery data for management purposes.
2) The accomodation of fishery extreme catch events (ECEs) has often been a neglected issue in the fishery modeling community. In light of our results, ECEs may play an important role in the correct estimation of fish abundance. Removing or windsorizing ECEs may result in an underestimation of abundance. This may imply that certain species' stock assessments have underestimated abundance. Despite the short term negative economic impact of this finding, driven by the modeling cost of this model, it will as well have a long term positive economic impact on the fishery.
3) The analysis of telemetry data using LGCPs to infer species habitat use will reduce modelling cost as it simplifies the required steps to model this sort of data. This may as well be applied to fishery distribution data, i.e. fishing locations, to account for the preferential sampling bias problem when estimating fish abundance using fishery dependent data only.
4) The analysis of acoustic data using Marked Point Processes, despite being work in progress, is an idea that was seeded within MultiSeaSpace. This new way of modeling fishery acoustic data provides a unique and statisticaly sounf tool to provide a probabilistic estimates abundance in this surveys. It is hard to estimate the economic impact of this for us, but if it works, we expect it to be an inflexion point in the estimation of uncertainty in fishery acoustic surveys, which will probably imply better fit in stock assessment models and therefore the management of the fishery.